Mobility Edge of L\'evy Matrices - Archive ouverte HAL Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2022

Mobility Edge of L\'evy Matrices

Amol Aggarwal
  • Fonction : Auteur
Patrick Lopatto
  • Fonction : Auteur

Résumé

L\'evy matrices are symmetric random matrices whose entry distributions lie in the domain of attraction of an $\alpha$-stable law. For $\alpha < 1$, predictions from the physics literature suggest that high-dimensional L\'{e}vy matrices should display the following phase transition at a point $E_{\mathrm{mob}}$. Eigenvectors corresponding to eigenvalues in $(-E_{\mathrm{mob}},E_{\mathrm{mob}})$ should be delocalized, while eigenvectors corresponding to eigenvalues outside of this interval should be localized. Further, $E_{\mathrm{mob}}$ is given by the (presumably unique) positive solution to $\lambda(E,\alpha) =1$, where $\lambda$ is an explicit function of $E$ and $\alpha$. We prove the following results about high-dimensional L\'{e}vy matrices. (1) If $\lambda(E,\alpha) > 1$ then eigenvectors with eigenvalues near $E$ are delocalized. (2) If $E$ is in the connected components of the set $\big\{ x : \lambda(x,\alpha) < 1 \big\}$ containing $\pm \infty$, then eigenvectors with eigenvalues near $E$ are localized. (3) For $\alpha$ sufficiently near $0$ or $1$, there is a unique positive solution $E = E_{\mathrm{mob}}$ to $\lambda(E,\alpha) = 1$, demonstrating the existence of a (unique) phase transition. (a) If $\alpha$ is close to $0$, then $E_{\mathrm{mob}}$ scales approximately as $|\log \alpha|^{-2/\alpha}$. (b) If $\alpha$ is close to $1$, then $E_{\mathrm{mob}}$ scales as $(1-\alpha)^{-1}$. Our proofs proceed through an analysis of the local weak limit of a L\'{e}vy matrix, given by a certain infinite-dimensional, heavy-tailed operator on the Poisson weighted infinite tree.

Dates et versions

hal-03866776 , version 1 (22-11-2022)

Identifiants

Citer

Amol Aggarwal, Charles Bordenave, Patrick Lopatto. Mobility Edge of L\'evy Matrices. 2022. ⟨hal-03866776⟩
15 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More